Estimation of DAS microseismic source mechanisms using unsupervised deep learning

Matthew Eaid, Hu Chaoshun, Lin Zhang, Scott Keating, Kristopher A. Innanen

Distributed acoustic sensing (DAS) is an increasingly prevalent technology for seismic acquisition, especially in reservoir monitoring settings. Recently, it has attracted interest for microseismic monitoring during hydraulic fracturing and as a complement to broadband seismometers for measuring teleseismic waves generated by earthquakes. A key component of these data is the source mechanism information encoded in the direct arrivals and how to best utilize this data to make inferences about the source mechanism, given DAS measurements of the direct arrivals is an open question. DAS is a relatively new technology, providing very large datasets of different physical aspects of the wavefield than geophones or seismometers. Consequently, conventional moment tensor inversion is challenging to transfer directly to DAS data. Instead, we turn our attention to deep learning algorithms for estimating these source mechanisms. Viewing our seismic data as containing diagnostic features of the source mechanism, we reduce the data to only those features most pertinent to moment tensor inversion by training a convolutional auto-encoder. The extracted features are then analyzed using clustering and generative adversarial networks (GAN). Clustering based on source mechanism is observed, confirming the extracted features contain important source mechanism information. Furthermore, we develop a trained GAN that provides an accurate mapping from feature space to moment tensor estimate, which shows promise when applied to a field DAS-microseismic dataset collected during hydraulic fracturing. Data modeled with the predicted source mechanism shows a strong correlation to the field data event.